Explainable and fair skin lesion diagnosis via ABCDE-grounded multimodal prompt learning

Deep learning has significantly advanced automated skin cancer diagnosis. However, most existing approaches rely solely on dermoscopic images and lack sufficient clinical context and interpretability. Although these methods often achieve dermatologist level classification accuracy, their limited reasoning transparency restricts clinical trust and broader adoption. This study proposes an explainable multimodal vision-language framework that integrates dermoscopic imagery, structured patient metadata, and ABCDE clinical rule-based dermatological prompts for melanoma-oriented multiclass skin lesion classification and clinically grounded explanation generation. The proposed framework was evaluated on the ISIC 2019 and PAD-UFES-20 datasets, achieving an average accuracy of \(92.4 ± 0.6\) %. Unlike prior post hoc explanation meth ods, the proposed framework generates grounded visual and text based clinical rationales aligned with skin lesion morphology and ABCDE dermatological cues. Experimental results demonstrate that the framework improves interpretability and transparency by generating clinically aligned model explanations and dermatological reasoning.